Physics informed neural networks for fusion plasma research

The electron density in the edge region of magnetically confined fusion plasma can be reconstructed from experimental data, e.g. alkali beam emission spectroscopy data. The traditional Bayesian approach is computationally expensive and can not be integrated into real-time control systems. We propose shifting the heavy computations to the training phase of neural networks and infer densities quickly, meeting real-time conditions. We introduce a physics informed loss function for regularization.